Cresta launches Conductor, a CX-focused agent builder on the Claude Agent SDK

Cresta has launched Conductor, a natural language agent builder designed to help teams create production-ready customer experience (CX) agents. Built on the Claude Agent SDK, Conductor acts as a meta-agent that guides developers from blueprint to implementation, evaluation, and optimization.

From Internal Tool to Product

Cresta originally developed an early version of Conductor using Claude Sonnet and Claude Opus to support internal customer deployments. The company sought to productize this capability, allowing customers to build agents directly within the platform. By using the Claude Agent SDK as a general-purpose execution harness, Conductor can gather context, use tools, write code, and adapt based on results, while Cresta layers its specific CX workflows and domain expertise on top.

Key Features and Architecture

Conductor is designed to address the complexity of enterprise AI agents, which must handle nuanced customer interactions like refund requests or billing issues. Key aspects of its architecture include:

• Knowledge Flywheel: As teams build agents, Conductor captures proven patterns and workflows as reusable memory artifacts, making future development faster and more consistent.
• Deterministic vs. Flexible Balance: The tool helps developers identify where to enforce strict business rules (for regulated or critical workflows) and where to allow model flexibility.
• Evaluation Module: Conductor turns requirements and edge cases captured during the build phase into automated tests, allowing teams to monitor workflows as business rules change.

Evaluation and Impact

Cresta evaluates Conductor based on outcome, execution path, quality, and resource use. The team runs specific build tasks—such as creating new agents, writing test cases, and performing root cause analysis—to measure performance. In early use cases across Cresta and its partners, the company reports that Conductor cut initial deployment time roughly in half.

The approach emphasizes that the initial build is only about 20% of the effort, with the remaining 80% dedicated to testing, optimization, and continuous improvement. Conductor is designed to support this iteration loop, ensuring agents remain reliable as they evolve.